How to use phonedriver with Cursor and other MCP clients

Add phonedriver to Cursor or any MCP client that starts a stdio server, with a JSON entry, an optional device setting and the list of tools it exposes.

Any client that starts a stdio MCP server can use phonedriver. I test with Claude Code. I have not run a Cursor session myself, so the Cursor steps below use the standard stdio entry and Cursor's documented configuration file, not something I watched work.

The entry

{
  "mcpServers": {
    "phonedriver": {
      "command": "npx",
      "args": ["-y", "phonedriver", "mcp"],
      "env": { "PHONEDRIVER_DEVICE": "android:emulator-5554" }
    }
  }
}

The first start downloads the binary once. On Windows, some clients cannot start npx directly: use "command": "cmd" and "args": ["/c", "npx", "-y", "phonedriver", "mcp"]. If you installed phonedriver yourself, use "command": "phonedriver" with "args": ["mcp"], or the full path. PHONEDRIVER_DEVICE is optional when one device is connected.

Cursor

Put the entry in ~/.cursor/mcp.json for every project, or in .cursor/mcp.json inside one project. Open Cursor's MCP settings and check that phonedriver shows its tools.

Other clients

Use the client's own place for stdio servers, with the same command and args. If the client cannot start a local process, run phonedriver where the client can reach it and forward the daemon's socket (see Android emulator).

Tools

The server has fourteen tools: devices, apps, snapshot, screenshot, tap, type, swipe, scroll, press, launch, wait, batch, clipboard and shell. apps also installs, stops, clears and uninstalls. Each prints the same text as the CLI verb of the same name, so anything you read in these docs about the command line holds for the tools.

In Claude Code's wrapper the fourteen tools cost about 51 tokens each whatever their text. The first turn of a one-word claude -p costs 5,606 tokens with phonedriver attached (token benchmark).

Approval prompts need elicitation

The approval guard asks you through MCP elicitation. A client that does not support elicitation cannot show the question, and the tool then replies that the tap needs the human's approval and taps nothing. Check that your client supports elicitation before you give an agent a phone with a shopping app on it.

PreviousClaude Code NextAndroid emulator